Field of the Invention
The present invention relates to a technique of handing information about an image or a region in an image.
Description of the Related Art
Image region division processing is widely used as preprocessing of object detection, scene analysis, and region-specific image processing (for example, noise reduction processing) because it can grasp the structural feature of an image by putting similar pixels in the image together and dividing the image into regions. Note that the image region division processing is also called segmentation or superpixel division. As a method of image region division, there are proposed various methods such as a method using clustering, mean shift, region growing, split-and-merge, and a graph-based method. In particular, a method using k-means clustering described in patent literature 1 (Japanese Patent Laid-Open No. 8-30787) is suitable for high-speed execution because of a light processing load. To further speed up the method using k-means clustering, SLIC described in non-patent literature 1 (Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk, “SLIC Superpixels”, EPFL Technical Report, 2010) or non-patent literature 2 (Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk, “SLIC Superpixels Compared to State-of-the-Art Superpixel Methods”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 34, num. 11, pp. 2274-2282, 2012) or gSLIC described in non-patent literature 3 (Carl Yuheng Ren and Ian Reid, “gSLIC: a real-time implementation of SLIC superpixel segmentation”, Technical report, Dept of Engineering Science, University of Oxford, 2011) is usable.
On the other hand, a label map shown in FIG. 1A is used as a method of expressing a region division processing result. The label map stores a region label (normally, a number) used to identify a region to which a pixel belongs on a pixel basis and needs data storage areas as many as the number of pixels. To reduce the data area to store the label map, a method of JBIG-compressing the label map into a bit plane, as described in patent literature 2 (Japanese Patent Laid-Open No. 9-98295) or a method of compressing the label map by combining reduction and binary image encoding, as described in patent literature 3 (Japanese Patent Laid-Open No. 10-4549) is usable.
In an image transmission system formed from a network camera and a server, when transmitting a result of region division processing executed in the network camera to the server and using it, the problem is what kind of format should be used to transmit the region division processing result. If data is transmitted in the format of the label map shown in FIG. 1A, data transfer as much as the image size is necessary, and a very large transmission band is used. When the label map compression technique described in patent literature 2 or 3 is used, the transmission data amount can be suppressed to a fraction of several tens. However, the transmission band is desired to be further decreased as much as possible.
When performing region division in the camera and performing processing using the result as well, the size of the memory to hold the region division result and the memory transfer band to read out the region division result and use it are problematic. In addition, when storing a region division result together with an image in an apparatus for storing image data, the size of the storage area used for storage similarly poses a problem.